Folding Forum

Foliant wrote:

PantherX wrote:

Foliant wrote:...can someone explain what the "Average performance" in "ns/day" is meaning?...

[...]
Here's an image which shows the various time-scales to put things in preservative:
https://www.frontiersin.org/files/Artic ... 8-g001.jpg
Source -> https://www.frontiersin.org/articles/10 ... 00258/full
[...]

That destroyed my thoughts about the speed of research that happens.
There is the need for more Hardware.

FYI I managed to get a 1050 (should be close but below the P106-090) for testing.
Tested under Windows because i dont know how to track PCIe-Bus-Load under Linux.

Code: Select all

PCIe 3 16x
Project: 16920 (5,30,62)
TPF: 1,54
Bus: 46%
Average performance: 75.6567 ns/day


PCIe 3 16x
Project: 16920 (26,15,74)
TPF: 1,54
Bus: 46%
Average performance: 75.2613 ns/day

Code: Select all

PCIe 2 1x
Project: 13422 (6078,69,1)
TPF: 7,45
Bus: 73%
Average performance: 36.7973 ns/day

PCIe 2 1x
Project: 13424 (1822,11,0)
TPF: 7,36
Bus: 71%
Average performance: 18.9391 ns/day

So maybe a 1060 / P106 is max for a PCIe2 1x Slot

Regards,
Patrick

Pcie bus load can be seen in a gui, in nvidia-X server, where also your fan speed and gpu frequency is displayed.
From your numbers, that 1050 is already operating at 50% performance on a 2.0 x1 slot. The performance of a faster gpu will only run at an even lower percentage.
A pcie bus load of above 60-75% is pretty much bottlenecked.
It means that for 60-75% of the time, the pcie bus is utilized above 50%, with plenty of peaks of 100% utilization.
While for games or other applications, 75% loads are occasional (eg: the changing of a scenery, the loading of new textures), for fah it matters a lot more. FAH accesses the pcie hundreds to thousands of times per wu, and each time you'll lose some PPD, as the gpu waits for all instructions to be loaded. The longer it takes to load, the more gpu cycles you lose in time.

You can also see loading times in the gpu subsection of the windows task manager. If your gpu (compute 0) has slight dips in it's 90-100% load, it means it's waiting for instructions, thus most likely pcie bottlenecked.

For running GPU tests on pcie 2.0 slots, Linux is really recommended,because it has less pcie overhead, and chances are a 1050 can still operate relatively well.
My estimation is, that since a 2060 needs a 2.0 x8 slot or higher in linux, a 1060 3GBrunning roughly at half the speed, I guess would need a 2.0 x4 slot, and a 1050 an x2 slot.

I personally believe, that pcie 2.0 x1 speeds are a bottleneck for most modern day computers. They may not bottleneck a pc running 1,6Ghz, with a GT 1030, but even if your gpu is this speed or slower, your cpu will more than likely be 50% (or more) faster.
And thus guarantees a slowdown at the pcie lanes for Fah.
FAH being largely dependent on the speedy completion of WUs.

So while it's entirely possible to use pcie 2.0 x1 slots for folding, you'll be wasting a lot of power, run the system very inefficient, and hold back the GPUs from running at their fullest potential, as they're waiting for instructions from the CPU.

Like said, pcie 2.0 x8 or below, really benefits from Linux. Especially if running faster GPUs.

PantherX wrote:Thanks for the information MeeLee & foldy, I will keep this in mind when I make my dedicated build in the future

I'd think you'll have a hard time finding any cpu with a lower than 3.5Ghz clock speed nowadays. Core count as well, it's even hard to find a quad core cpu.
Most cpus sold now start at 6 threads.

Not thing you'd have to take in consideration when building a new rig, is that big navi (6000xt), rtx 3070-3090 will need a cpu running at above 3Ghz.
This makes running 2 GPUs per core not a good idea.
Then again, at 225-350W per GPU, as well as the pcie 4.0 x8/x8/x4 lanes limitation, it may be very hard to run more than 3 GPUs per system, without thermally throttling.

If I were to make a new system today, I'd get a ryzen 3 with a pair of RTX 3070 GPUs (very close to 2x 2080Ti).
If I'd want to buy a new system in a year, I'd probably get 2x RTX 3090 gpus for it instead. I don't think we'll see many WUs that will fully utilize the 3090 yet...

MeeLee wrote:...I don't think we'll see many WUs that will fully utilize the 3090 yet...

There are two potential way to utilize those high-end GPUs:
1) Simulation with Millions of atoms
2) Running multiple WUs on a single GPU

Development is aware of those ideas (and potentially others too) and it is something that they are thinking about. There's no plan or ETA but this is an evolving situation so let's hope for the best and see what happens